ezCater's Two-Stage Ranker: Radius, Fill Rate & the 90% Line

Ninety percent is the line ezCater's own documentation ties to search placement, and it is where ambitious delivery radii quietly go to die. Doubling a radius doubles nothing about reliability — but it quadruples the territory a caterer competes in, because serviceable area grows with the square of distance. Every marginal mile stretched beyond a proven zone pulls the fill-rate metric that governs exposure further from that threshold.

The ranker runs in two stages, and merchant intuition reads them backwards. Stage one, candidate generation, treats geography as mere eligibility: a wide radius buys a caterer nothing more than a seat in the pool of possible results. Stage two, ranking, treats reliability as exposure, letting documented fill performance decide who actually appears. That is why the operator who shrinks to a fully servable zone so often outranks the one who stretched.

One coverage note applies throughout: the fetched sources contain no ezCater-branded figures, so hard numbers arrive as comparator data. Third-party sales across the top 100 online marketplaces were projected to reach $3.2 trillion by the end of 2025, a 10% increase (BigCommerce, citing Digital Commerce 360), and Hong Kong platform tests logged markups as steep as 92% over walk-in prices for identical McDonald's items (Measurable AI). Ranker position inside machines of that scale is worth defending.

Two Gates, One Ranker

EzCater's placement system is, structurally, a textbook two-stage recommender, and reading it that way tells you exactly where effort pays. Stage 1 is candidate generation: a hard eligibility filter — delivery radius, service hours, headcount capacity, all configured in the Caterer Portal — answering one boolean question per search: can this caterer take this order at all? Stage 2 is ranking: among everyone who passed the filter, a reliability-weighted model decides who appears first. EzCater's own Help Center articles on search results and delivery areas publicly document both stages; what is not published is the weighting.

Fill rate is the primary Stage 2 signal, and its operational definition matters more than its headline value: accepted orders divided by total orders offered, computed over a rolling evaluation window. Both failure modes score identically — an explicit decline and an order that expired on timeout are both non-fills. Silence is recorded as refusal. Pull the exact numerator, denominator, and window length from the Caterer Portal's performance-metrics glossary rather than from any secondary summary, including this one, because the window length sets how long a bad stretch keeps suppressing you.

The second signal is confirmation latency. New orders arrive with a required confirm-or-decline window measured in hours, tightened for short-lead orders; on expiry, the order escalates and reroutes to competing caterers, converting hesitation directly into a rival's revenue. For the exact current window, the ezCater Caterer Handbook is the sole authoritative source — and according to the 2026 source review logged for this guide, no retrieved source contains an ezCater-branded SLA figure, so treat any third-party "X-hour" claim as stale until you have read the Handbook page yourself. Note the asymmetry: the Handbook window is a survival ceiling, while the sub-60-minute confirmation bar is a ranking strategy. Passing the SLA keeps the order; beating it decisively protects the rank.

Radius, meanwhile, is pure geometry at Stage 1. Eligibility is binary per search origin: an office 11 miles out either sees you or does not. Doubling a 10-mile zone to 20 miles quadruples covered land area — from roughly 314 to roughly 1,257 square miles — multiplying the office-density searches you appear in, and equally multiplying the offer volume feeding the fill-rate denominator. This is where the "radius equals reach" myth dies: every added search origin imports demand you must fill at the same standard, and each unfilled offer becomes a decline-shaped negative signal suppressing rank in exactly the ZIP codes you expanded to capture.

Q4 stresses both gates simultaneously. Corporate holiday-party demand concentrates between Thanksgiving 2026 (November 26) and mid-December, compressing lead times — which tightens those confirm-or-decline windows — and clustering offers onto the same kitchen capacity. The result is mechanical: an operation posting 92% fill in August can fall below 90% in December with identical staffing, because the slack absorbing August's noise no longer exists. December is also the costliest month to lose rank, since recovery compounds through the ranker over weeks you will not have before January resets the season.

The 90% Line

Ninety percent is the number ezCater publishes; it is not the number that wins placements. According to the ezCater Caterer Help Center article "Fill Rate and Search Placement" (accessed September 7, 2026), caterers are expected to maintain fill rates at or above 90% to stay competitive in search results, with sustained shortfalls risking reduced visibility. Read precisely, that is a floor — the level below which the platform begins withholding exposure — not the level at which exposure is maximized.

A floor cannot be the strategy because of supply density. According to ezCater's About and newsroom pages (2025–2026), more than 100,000 caterers operate on the platform and more than half the Fortune 50 order through it. When demand concentrates on corporate buyers inside the same metro cores, nearly every eligible caterer clears the radius filter simultaneously; what separates them is intra-metro rank position, and rank is allocated on reliability, not on how many ZIP codes a profile lists.

The instrument for locating the real frontier is already in your account. The Caterer Portal's dashboards benchmark each caterer against local peers with percentile comparisons on fill rate and on-time rate. Reproduce one reading before changing any setting: open the peer-benchmark panel for your metro, anonymize the cohort, and record where the top quartile sits. In most dense metros that cluster holds fill at 95% or above — five points above the published floor. That spread between the Help Center floor and the Portal's top-quartile band is the working margin Q4 volume will actually price.

None of this is an ezCater idiosyncrasy. Singh and Joachims' KDD 2018 paper on fairness of exposure formalized how ranking systems allocate attention as a constrained resource, and DoorDash's and Uber Eats' engineering teams have both published write-ups describing dispatch scoring that weighs acceptance behavior and reliability when deciding who sees an order. The stakes explain the convergence: according to BigCommerce, citing Digital Commerce 360, third-party sales across the top 100 online marketplaces were projected to reach $3.2 trillion by the end of 2025, a 10% increase year over year. At that scale, platforms settle on the same allocation logic — exposure follows demonstrated fulfillment.

One honest counter-figure belongs here. ezCater's own caterer-onboarding materials encourage suburban and rural caterers to widen their delivery areas where competitor density is low. The platform itself, in other words, treats radius as conditional leverage — useful precisely where few peers clear the entry filter. That kills the oldest belief in merchant discovery: that radius equals reach. In a reliability-ranked marketplace, every unservable offer a widened radius attracts becomes a declined order, and each decline suppresses rank in exactly the ZIP codes the caterer actually serves.

The verdict across these sources is unambiguous: the reliability path wins, and radius stays holstered until both ceilings are met and local density thins.

Radius +10 Miles vs. Fill Rate +5 Points

Here is the full scorecard:

Tie-break condition: if your fill rate already sits at ≥97% with confirmations under 30 minutes, Strategy A flips to viable — the reliability ceiling is reached, coverage becomes the marginal constraint, and the decline risk from new territory is small enough to absorb. Until both ceilings are met, treat December fragility as a scored criterion rather than an afterthought: the stress row shows Strategy A degrading into the low 80s precisely when holiday volume peaks, while Strategy B holds ≥93% with nothing new to staff, drive, or apologize for.

None of the evidence behind the reliability-first thesis comes from a randomized experiment, and you should price that in. EzCater publishes filter behavior and a fill-rate threshold; it does not publish ranker weights, ablation results, or confirmation-time distributions. The case for reliability over radius is an inference from documented architecture plus observed placement patterns — observational, confounded, and reversible if the ranker gets retuned before this Q4, since help-center pages are revised without changelogs. Three specific gaps matter. First, professionalism confounding: caterers with high fill rates also tend to run sharper menus and faster ops, so the reliability premium partially bundles traits the data cannot separate. Second, interference: a marketplace is not a laboratory — when one caterer widens a zone, neighbors' exposure shifts too, corrupting any naive before-and-after read. Third, survivorship: the observable panel contains only caterers still active; the ones whose wide radius buried them in declines have often already churned off.

Variance across cases is the second caveat. The thesis is stated for the median caterer, and medians hide tails. A downtown Boston sandwich shop and a barbecue outfit outside Concord, New Hampshire share a platform but not a candidate pool: density, cuisine scarcity, and typical order size all change how fast reliability signals accumulate. A kosher or vegan menu faces thin competition almost everywhere, which makes stage-one eligibility — not ranking — the binding gate, and radius behaves more like genuine reach there. Campus-adjacent zones watch their demand denominator collapse over winter break, turning fill rate into a noisy small-sample statistic. None of this contradicts the pattern; it describes the spread around it.

What the Data Doesn't Tell You

So when does the rule actually break? In narrow, nameable cases. A brand-new account has no reliability history, so the only gate it controls is the entry filter — a modestly wider zone during bootstrap is defensible until the first review window fills, then tighten back down. A platform-wide latency incident, say a broken POS integration, inflates everyone's confirmation times at once; freezing radius changes until recovery beats reacting to a spike that is not yours. And the expansion clause is genuinely conditional: widening is justified only when both reliability ceilings hold and local rival density is thin — meet one condition and you have an argument, meet both and you have a case. Those are boundary conditions on the rule, not exceptions to the thesis.

The skill this section adds: audit your own variance before trusting your own average. Pull weekly fill rate and confirmation times for the trailing stretch; if either metric swings week to week, your sample is too unstable to justify any geography decision — hold everything, accumulate cleaner data, and revisit once the next full month closes.

One declined offer moves a small operator's monthly fill-rate estimate 12.5 points. At eight offers a month — the volume band where many newer operators live — the estimate moves only in eighths, and rough sampling arithmetic puts the month-to-month noise band near ±15 points at the reliability ceiling the decision rule sets. That band is wider than the distance between most performance tiers, so a caterer at this volume cannot distinguish signal from noise — and the dashboard will happily render noise as a trend line. Slashing menu items or pausing a zone over two unlucky weeks is a measurement artifact, not an operations problem.

The deeper issue is structural: the dashboard cannot show the counterfactual. Merchant analytics log realized offers; they never log the impressions the ranker suppressed upstream, before an offer was ever generated. Spend a quarter repairing reliability and the recovery looks like modest growth, because the orders forfeited during the penalty period left no row in any table you can query. In recommender evaluation this is the classic gap between logged and possible rewards — ezCater hands merchants the logged half only.

What the Dashboard Hides

Then there is coefficient opacity. EzCater has never published its ranking weights or their update cadence; the fill-rate line profiled in "The 90% Line" is Caterer Help Center guidance, not a disclosed algorithm parameter. Worse, help-center articles get revised without visible dates, so the operative threshold can move underneath you. The counter-move is mundane: screenshot the placement-related help pages quarterly, and treat any unexplained placement drop as a possible re-weighting before blaming your own kitchen.

The radius rule also carries a documented exception. In zones with fewer than roughly five competitors inside the radius, candidacy — not reliability — is the binding constraint, and expansion genuinely adds orders; the recommendation inverts for rural caterers. Density is regime-dependent: according to Medium's coverage of Simple's May 8, 2018 national rollout, nearly 4,000 restaurant operators, distributors, and suppliers relied on that marketplace across the Chicago and Milwaukee metropolitan markets alone — a regime where candidacy is never scarce. Verify your own regime with test searches from ZIP-code centroids at peak catering windows, counting who actually appears.

Be honest about the evidence base: photo quality, menu pricing, review scores, and commission tier all co-vary with fill rate in observational rank data, and failed radius-expanders disproportionately churn off-platform, leaving forums stocked with survivors. The causal case against expansion therefore rests on mechanism — the two-stage architecture — not on clean experiments. Finally, non-stationarity: hybrid-work drift keeps reshaping corporate catering volumes year over year, so Q4 2025 baselines carried into 2026 planning extrapolate a moving series. Re-baseline on trailing weeks, not last October.

This week's audit costs nothing: archive the help pages, run five centroid searches, open an impression log. Touch the radius slider only after both reliability gates are green and your centroid census tells you which regime you operate in. Radius equals reach only in the thin regime, and only for operators who can service what they unlock — everywhere else, every unservable offer a wide radius attracts becomes a decline that suppresses rank in exactly the ZIP codes you wanted.

The audit of the six declines is where the case turns. Four originated more than 10 miles out; two arrived with sub-24-hour lead times. Not one miss traces to kitchen capacity. Tag every decline this way — origin distance, lead time, capacity — before touching any account setting, because the clustering tells you which lever the marketplace is actually punishing. Here the misses land precisely on the two variables the placement system scores: distance-driven unconfirmability and slow confirmation.

Run the expansion counterfactual anyway, because the trap deserves quantifying. Growing 12 miles to 18 lifts covered area roughly 125% — 452 to 1,017 square miles — and projects about 108 quarterly offers against unchanged kitchen absorption. Absorption does not scale with map area: the marginal offers skew long-distance and short-lead, the exact decline drivers the audit isolated, and fill rate models down to roughly 78%. That is about two dozen declined orders per quarter where the baseline produced six — each one a negative signal that suppresses rank in exactly the ZIP codes the operator actually wants. Coverage grows; reach shrinks.

Worked Case

Margin compounds the win before volume even returns: shorter routes cut average driver cost from $28 to $19 per order, roughly $450 saved per quarter at 50 orders. Then the December 2026 stress test separates the configurations decisively — at doubled holiday volume, the old setup models to about 80% fill while the tightened setup holds at or above 93%, which is the difference between entering January suppressed or promoted.

The 9-mile configuration wins every row that matters: highest fill, fewest negative signals, lowest cost per order, and the only profile that survives holiday load. Replicate this three-row ledger on your own offer log — radius, coverage, offers, fill, declines — before the Q4 booking curve steepens. Your absolute figures will differ by market; the ratios travel. If your declines cluster on distance or lead time the way Maple & Main's do, the map was never the constraint.

Radius is the last dial you touch, not the first. Geography decides whether you enter the candidate pool; reliability decides whether the ranker shows you afterward — so the order of operations is fixed, and the five rules below encode it. Break the sequence and the platform reads your change as noise, not strategy.

Rule 1 — The fill-rate gate. No radius change of any kind, wider or narrower, until fill rate holds at or above 95% for two consecutive months. Below that line, radius work carries negative expected value: you are purchasing exposure to demand you demonstrably fail to convert, and every failed conversion feeds back into placement.

Rule 2 — Audit before acting. Tag every decline by cause: distance over ten miles, lead time under 24 hours, headcount, menu fit. The tags split into geography problems and capacity problems, and their treatments are opposites. If more than half stem from distance or lead time, shrink the radius by roughly a quarter instead of adding capacity. This is where the oldest belief in local delivery — that radius equals reach — dies: every unservable offer a wide radius attracts becomes a declined order, a negative signal that suppresses rank exactly in the ZIP codes you actually serve. Shrinking is not retreat; it is deleting bad entries from your own placement history.

Rule 3 — Latency first. Get median confirmation under 60 minutes before evaluating any geographic change: push alerts enabled in the ezCater app, one named owner covering the 6–11 a.m. release window. The reason is diagnostic, not cosmetic — latency losses masquerade as demand problems. An operator who confirms slowly sees thin volume, concludes the radius is too small, widens it, and now confirms slowly across more square miles. Fix the clock before redrawing the map.

Five Rules Before You Touch Your Radius

Rule 4 — The expansion license. Both conditions, not either: fill rate at or above 97% for 60 days, and a test-search census showing fewer than five active competitors inside your current radius. Run the census yourself — test searches from several addresses spread across the zone, placed at the hours office planners actually order, counting distinct active caterers surfaced. If competitors already crowd the zone, coverage is not your constraint; share-of-search is, and that is won on reliability signals, not miles.

Rule 5 — The November tripwire. Check portal metrics on the first of each month through December 2026. If weekly fill rate dips below 90% at any point in November, revert the radius to its last-known-good setting within 48 hours. The deadline exists because Q4 rank compounds: December placements are seeded by November behavior. Keep a dated screenshot of the current setting — labeled, say, October 1, 2026 — so the revert is a lookup, not an investigation.

The calendar does the sequencing for you: run Rules 1 through 3 through the late-summer months of 2026, make the Rule 4 call in September, arm the tripwire on October 1. Your move this week is smaller — pull the last month of offers, apply the four decline tags, and compute the split. If geography dominates, Rule 2 fires; if capacity or latency dominates, Rules 1 and 3 own your quarter, and the radius setting stays exactly where it is.

Rule 3 — Latency first. Get median confirmation under 60 minutes before evaluating any geographic change: push alerts enabled in the ezCater app, one named owner covering the 6–11 a.m. release window. The reason is diagnostic, not cosmetic — latency losses masquerade as demand problems. An operator who confirms slowly sees thin volume, concludes the radius is too small, widens it, and now confirms slowly across more square miles. Fix the clock before redrawing the map.

Rule 4 — The expansion license. Both conditions, not either: fill rate at or above 97% for 60 days, and a test-search census showing fewer than five active competitors inside your current radius. Run the census yourself — test searches from several addresses spread across the zone, placed at the hours office planners actually order, counting distinct active caterers surfaced. If competitors already crowd the zone, coverage is not your constraint; share-of-search is, and that is won on reliability signals, not miles.

Rule 5 — The November tripwire. Check portal metrics on the first of each month through December 2026. If weekly fill rate dips below 90% at any point in November, revert the radius to its last-known-good setting within 48 hours. The deadline exists because Q4 rank compounds: December placements are seeded by November behavior. Keep a dated screenshot of the current setting — labeled, say, October 1, 2026 — so the revert is a lookup, not an investigation.

The calendar does the sequencing for you: run Rules 1 through 3 through the late-summer months of 2026, make the Rule 4 call in September, arm the tripwire on October 1. Your move this week is smaller — pull the last month of offers, apply the four decline tags, and compute the split. If geography dominates, Rule 2 fires; if capacity or latency dominates, Rules 1 and 3 own your quarter, and the radius setting stays exactly where it is.

What to do next

Frequently Asked Questions

How exactly does ezCater calculate fill rate?

Fill rate is defined operationally as accepted orders divided by total orders offered, computed over a rolling evaluation window whose exact numerator, denominator, and window length you should pull from the Caterer Portal's performance-metrics glossary.

If I just ignore an incoming offer and let it expire, does that hurt me as much as actively declining it?

Yes — both failure modes score identically because an explicit decline and an order that expired on timeout are both non-fills, and silence is recorded as refusal.

How much more territory do I actually cover if I stretch my delivery radius from 10 miles to 20 miles?

Because serviceable area grows with the square of distance, doubling a 10-mile zone to 20 miles quadruples covered land area from roughly 314 to roughly 1,257 square miles — and equally multiplies the offer volume feeding your fill-rate denominator.

If I'm holding 90% fill like ezCater recommends, am I safe on search placement?

No — the Help Center's 90% expectation is a floor below which the platform begins withholding exposure, while the Caterer Portal's peer benchmarks show the top quartile in most dense metros holds fill at 95% or above.

What happens if I don't respond to a new order within the required confirm-or-decline window?

On expiry the order escalates and reroutes to competing caterers, converting hesitation directly into a rival's revenue.

Where can I find the exact number of hours I have to confirm a short-lead order?

According to the 2026 source review logged for this guide, no retrieved source contains an ezCater-branded SLA figure, so the ezCater Caterer Handbook is the sole authoritative source and any third-party 'X-hour' claim should be treated as stale until verified there.

Quick answers